CN-120147868-B - Movable landslide identification method

CN120147868BCN 120147868 BCN120147868 BCN 120147868BCN-120147868-B

Abstract

The invention discloses an active landslide identification method, which relates to the field of disaster identification and comprises the steps of obtaining image data of an area to be identified, constructing an active landslide data set, constructing a context-aware adaptive fusion model, wherein the context-aware adaptive fusion model comprises an encoder and a decoder, the encoder comprises a plurality of characteristic processing layers and a convolution layer which are sequentially connected, the characteristic processing layers comprise a convolution layer and a wavelet transform downsampling WDB module which are sequentially connected, the decoder comprises a plurality of multi-branch scale adaptive aggregation MSA modules, a plurality of characteristic extraction refinement layers and a convolution layer which are sequentially connected, the characteristic extraction refinement layers comprise a convolution upsampling CUB module which are sequentially connected, training the context-aware adaptive fusion model based on the active landslide data set, obtaining an active landslide identification model, and inputting the active landslide data set to be detected into the active landslide identification model to obtain an active landslide identification result.

Inventors

SONG CHUANG
LIU ZHENJIANG
HE KELU
WANG JIATONG
Nai Yihan
CAI XINGMIN
LI ZHENHONG
YU CHEN
CHEN YI
CHEN BO
ZHOU JIAWEI
WEN FAN
ZHANG XUESONG

Assignees

长安大学

Dates

Publication Date: 20260508
Application Date: 20250227

Claims (5)

1. A method of identifying an active landslide comprising: acquiring image data of a region to be identified, and constructing an active landslide data set; Constructing a context-aware adaptive fusion model, wherein the context-aware adaptive fusion model comprises an encoder and a decoder; the encoder comprises a plurality of characteristic processing layers and a convolution layer which are connected in sequence, wherein the characteristic processing layers comprise the convolution layer and a wavelet transform downsampling WDB module which are connected in sequence; The decoder comprises a plurality of multi-branch-scale self-adaptive aggregation MSA modules, a plurality of feature extraction refinement layers and a convolution layer, wherein the multi-branch-scale self-adaptive aggregation MSA modules are used for receiving feature inputs of two adjacent levels, adaptively distributing level weights according to target features and fusing, and the feature extraction refinement layers comprise the convolution layer which is sequentially connected, and a convolution up-sampling CUB module which is used for restoring resolution of features and refining boundary information; The multi-branch scale self-adaptive aggregation MSA module receives feature inputs of two adjacent levels, adaptively distributes level weights according to target features and fuses the feature inputs, and comprises the following steps: The method comprises the steps of performing feature enhancement on input low-level features by adopting 1X 1 convolution, and adjusting the channel number and spatial resolution of input high-level features by adopting the 1X 1 convolution and bilinear interpolation; Determining feature components according to the adjusted low-level features and high-level features, determining weights of the feature components by adopting a sigmoid function, splicing each group of weighted high-level features with the corresponding weighted low-level features to generate fusion features, carrying out normalization processing on each group of fusion features by adopting layer normalization, carrying out multi-scale feature extraction on the normalized four groups of fusion features by adopting expansion convolution, splicing the fusion features after feature extraction, and carrying out fusion feature output by adopting a 1X 1 convolution adjustment channel number; The convolution upsampling CUB module restores the resolution of the features and refines the boundary information, and comprises the following steps: The method comprises the steps of carrying out feature extraction and refinement on input features through a3×3 convolution layer, then carrying out 2 times up sampling on the input features, inputting two parallel depth separable convolution branches, extracting spatial features in horizontal and vertical directions of the input features by adopting convolution kernels of 1×3+3×1 and 3×1+1×3, carrying out feature fusion on outputs of the parallel branches by adopting a pixel-by-pixel addition mode, carrying out feature enhancement operation on the fused features by adopting a ReLU activation function, and finally adjusting the channel number of a feature map by using 1×1 convolution; Inputting the movable landslide data set into the context-aware adaptive fusion model for training to obtain a movable landslide recognition model for landslide recognition; and inputting the movable landslide data set to be detected into the movable landslide identification model to obtain a movable landslide identification result.
2. The method for identifying an active landslide of claim 1, wherein the acquiring image data of the area to be identified and constructing an active landslide data set specifically comprises: acquiring image data of an area to be identified by adopting an InSAR interferometry technology, and acquiring average deformation rate by utilizing INSAR STACKING technology according to the surface deformation image data; The average shape change rate value is mapped to an RGB color space, a raster image is determined, and a raster image sliding window is cut to generate a movable landslide data set for movable landslide identification.
3. The active landslide identification method of claim 1 wherein the downsampling of the wavelet transform downsampling WDB module comprises the steps of: The method comprises the steps of decomposing an input characteristic into four frequency domain characteristic components through Haar wavelet transformation, splicing the four frequency domain components along a channel dimension, and refining the characteristic representation through 1X 1 convolution, wherein the four frequency domain characteristic components comprise a low frequency component and three high frequency components.
4. The active landslide identification method of claim 1 further comprising: And carrying out post-treatment on the movable landslide identification result by using the average gradient value to obtain landslide connected domains, carrying out connected domain analysis on the landslide connected domains, extracting a gradient value set of each landslide connected domain in a corresponding gradient map, determining the average gradient of the connected domains, screening each landslide connected domain by setting a gradient threshold value, and recombining the screened connected domains to generate an optimized movable landslide identification result.
5. An active landslide identification method of claim 1 wherein the active landslide data set is filtered and denoised.

Description

Movable landslide identification method Technical Field The invention relates to the technical field of disaster identification, in particular to an active landslide identification method. Background Landslides, one of the common geological disasters, cause serious life threatening and property loss to people worldwide. Landslide events typically occur with slow active deformation phases, which may manifest themselves as the generation and propagation of surface soil and rock cracks, forming a wide active area over time, and eventually developing landslide instability. The active deformation area of the active landslide ALs is identified in time, so that the high-risk area is monitored and early-warned, and the method is very important for disaster prevention and reduction. The synthetic aperture radar interferometry InSAR technology can capture the tiny displacement of potential movable landslide, so that unstable and stable slopes can be distinguished in a large range, and the method is widely applied to the identification field of movable landslide ALs. The existing InSAR-based method for identifying the movable landslide ALs mainly comprises manual visual interpretation, cluster analysis and a threshold method, wherein the manual visual interpretation is time-consuming and laborious, and depends on expert knowledge, the adaptability of the cluster analysis to complex terrains is poor, the threshold method needs to depend on expert experience to determine a threshold value, the result is subjective, a low-contrast area is difficult to process, the identification precision of the movable landslide ALs is influenced, in addition, although the deep learning method can automatically identify the ALs in a large-scale terrains, the problems of small-scale landslide feature loss and poor boundary extraction effect still exist in the current research, and the problem of insufficient identification precision of the movable landslide ALs is caused. Disclosure of Invention The invention provides a method for identifying an active landslide, which is used for solving the problems in the prior art, namely how to improve the identification precision of ALs of the active landslide, and comprises the following steps: acquiring image data of a region to be identified, and constructing an active landslide data set; Constructing a context-aware adaptive fusion model, wherein the context-aware adaptive fusion model comprises an encoder and a decoder; the encoder comprises a plurality of characteristic processing layers and a convolution layer which are connected in sequence, wherein the characteristic processing layers comprise the convolution layer and a wavelet transform downsampling WDB module which are connected in sequence; The decoder comprises a plurality of multi-branch-scale self-adaptive aggregation MSA modules, a plurality of feature extraction refinement layers and a convolution layer, wherein the multi-branch-scale self-adaptive aggregation MSA modules are used for receiving feature inputs of two adjacent levels, adaptively distributing level weights according to target features and fusing, and the feature extraction refinement layers comprise the convolution layer which is sequentially connected, and a convolution up-sampling CUB module which is used for restoring resolution of features and refining boundary information; The multi-branch scale self-adaptive aggregation MSA module receives feature inputs of two adjacent levels, adaptively distributes level weights according to target features and fuses the feature inputs, and comprises the following steps: The method comprises the steps of performing feature enhancement on input low-level features by adopting 1X 1 convolution, and adjusting the channel number and spatial resolution of input high-level features by adopting the 1X 1 convolution and bilinear interpolation; Determining feature components according to the adjusted low-level features and high-level features, determining weights of the feature components by adopting a sigmoid function, splicing each group of weighted high-level features with the corresponding weighted low-level features to generate fusion features, carrying out normalization processing on each group of fusion features by adopting layer normalization, carrying out multi-scale feature extraction on the normalized four groups of fusion features by adopting expansion convolution, splicing the fusion features after feature extraction, and carrying out fusion feature output by adopting a 1X 1 convolution adjustment channel number. The convolution upsampling CUB module restores the resolution of the features and refines the boundary information, and comprises the following steps: The method comprises the steps of carrying out feature extraction and refinement on input features through a3×3 convolution layer, then carrying out 2 times up sampling on the input features, inputting two parallel depth separable convolution branches, extracting sp